Semantic Memory:

Rules of the Game

Naveen Kashyap, PhD

Indian Institute of Technology Guwahati
Email: naveen.kashyap@iitg.ernet.in
People hold on to information years after storing them

These contains knowledge relating to definitions of words,

arithmetic facts and procedures, historical, scientific and
geographical knowledge to name a few

The organization of the knowledge follows the metaphor

of a book shelf

Information in memory consists of knowledge for specific

events and memory for general knowledge

Endel Tulving (1972, 1983) argued that LTM contains two

stores namely episodic and semantic which although are
distinct but highly interactive.
Episodic – Semantic distinction

Episodic – enables people to travel back in time and

become consciously aware of witnessing or participating in
events in earlier time

Semantic – contains general and world knowledge,

arithmetic rules, past tense of verbs etc

Episodic Semantic
Personal Experience Facts and Concepts
Remember “When”… Remember “What”….
Temporally Organized Meaning related
Organization
 Semantic memory models
The Hierarchical Semantic Model
The model was proposed by Collins & Quillinan (1969). They
tested the idea that semantic memory is analogous to a network
of connected ideas.

The Model consists of nodes (in this case words/concepts). Each

node is connected to related nodes by means of pointers.

Thus the node that corresponds to a given word/concept

together with the pointers to other nodes to which the first node
is connected, constitutes the semantic memory for that
word/concept. The collection of nodes associated with all the
words & concepts is called Semantic Network
Collins & Quillian (1969) tested the principle of cognitive
economy with their model of semantic memory. They reasoned
that the closer a fact is stored to a particular node, the less time it
should take to verify the fact and property.

They reported that people took less time to respond to

sentences whose representations should span two levels (A
canary is a bird) than for those whose representation should
span three (A canary is an animal)

This model was called the hierarchical semantic network

model of semantic memory. The nodes in this model are
organized in hierarchy and most nodes have super ordinate
and subordinate nodes. Super ordinate nodes correspond to the
category name for which the thing corresponds to the subordinate
node was a member.
Meyer & Schvaneveldt (1971) reasoned that if related words
are stored close by one another and are connected to one
another in a semantic network, then when ever one node is
activated or energized, energy spreads to the related nodes.
One reason for such a fact could be the concept of spreading
activation, the idea that excitation spreads along the
connection of nodes in a semantic network
Limitations of HMSM
1) Cognitive Economy: Conrad (1972) found that people
respond no faster to sentences such as “A shark can move”
than to “A fish can move” or “An animal can move”.
2) Hierarchical Structure: Rips, Shoben & Smith (1973)
showed participants were faster to verify “A pig is an animal”
than to verify “A pig is a mammal” thus violating the
hierarchical structure (animal-mammal-pig)
3) Typicality Effect: Rips (1973) found that responses to
sentences such as “A robin is a bird” were faster than
responses to “A turkey is a bird”. In general typical instances
of a concept is responded to more quickly than atypical
instances.
 The Feature Comparison Model
Smith, Shoben & Rips (1974) proposed an alternative to the
HMSM called the feature comparison model of semantic
memory.
Assumption: The meaning of any word or concept
consists of a set of elements called features. Features
come in two types

1) Defining – meaning that the features must be present in

every example of the concept and

2) Characteristic – meaning the features is usually but not

necessarily present.
Feature Comparison model can explain shortcomings of the
HMSM
1) Typicality Effect: Sentences like “A robin is a bird” are
verified more quickly than “A turkey is a bird” because robin
being more typical examples of birds are thought to share
more characteristics feature with “bird” than do turkeys
2) Category Size Effect: The feature comparison model
assumes that as categories grow larger they also become
more abstract which lead to lesser defining features.
Criticisms:
1) There is no existence of defining features (suppose a bird
has clipped wings. Will it no longer be a bird?)
Other Network Models
Collins & Loftus (1975) presented an elaboration of the
Collins & Quillian (1969) Hierarchical model that was the
spreading activation model

This model conceives of semantic memory as a

network, with nodes in the network corresponding to
concepts. They also saw related concepts as connected
by paths in the network
They further asserted that when one node is activated the
excitation of the node spreads down the paths or links to
related nodes.

They believed that as activation spreads down the paths or

links to related nodes. When activation spreads outwards,
it decreases in strength, activating very related concepts a
great deal but activating distantly related nodes only a little
bit.
In this model very similar concepts have many connecting
links and are placed close to each other. Each
link/connection between two concepts is thought to have a
certain weight or set of weights associated with it.
Criticism: The breadth of the model makes it difficult to make
clear and strong predictions from the model regarding
empirical findings.
Anderson’s ACT model
Proposed by John Anderson (1976, 83, 93) and called the
adaptive control of thought model of memory (ACT, ACT-*,
ACT-R). Based on analogies to computers, ACT gives rise to
several computer simulations of cognitive processing of
different tasks.

Distinguishes among three types of memory systems:

– Declarative memory (information, facts)
– Working memory (information that is currently at
a high level of activation)
– Procedural memory
Declarative Memory
Anderson (1983) believed declarative memory
stores information in networks that contain
nodes.

There are different types of nodes, including those

corresponding to spatial images or to abstract
propositions.

ACT model allows both for activation of any node

and for spreading activation to connected nodes
Procedural Memory

Representation is as a series of “production rules.”

If–then statements that tell how to perform a

particular action

Production rules specify a goal to achieve, one or

more conditions that must be true for the rule to
apply, and one or more actions that result from
applying the rule
Connectionist models

Model “learns” to develop patterns of activation through

many trials with training examples (back propogation)

Initially, connections between nodes are set at random

strengths (weights); experience leads these connections
to be activated more or less strongly.

Training occurs by presenting a specific example to the

network which then generates a particular output.

Training takes places in ‘epochs’. Each epoch produces an

output activation which is compared with correct target
activation.
A connectionist model for “robin”
Schemata

Represents knowledge in Semantic Memory (Bartlett, 1932)

Schemata contains: General Knowledge about world and

Information about particular events

Schema – are large units of organized information which are

used for representing concepts, situations, events actions in
memory

Rumelhart and Ortony (1977) – fundamental building blocks

of cognition, units or organized knowledge analogous to
theories
Schemata – packets of information that contain both
variable and fixed part.
Mammal, four legs
DOG (fixed part)
Schemata
Breed, Size, Color, Temperament
(variable part)

Schemata – indicated relationship among various pieces

of information

Schemata – connected to other schemata in a variety of

ways
Schemata – fills in default values of certain aspects of
the situations which help us in making assumptions

Schemata – exists at all levels of abstraction; thus they

can exist for small parts of knowledge and for very
large parts

Scripts – are schema for routine events (for example

going to a restaurant)

Scripts – are used across variety of situations for

figuring out unknowns (new city visit)
Scripts – help us make a number of inferences

Scripts – help us for order.

Bower, Black and Turner (1979) showed that when

information from a story is presented in jumbled up
sequences people tended to recall the story in scripted
order.